Hospitals Follow Banks’ Playbook for Taming Generic AI

Physicians, nurses, researchers and staff will get secure access to an AI-powered clinical reasoning platform that synthesizes peer-reviewed medical studies and clinical guidelines, with citations attached to every response, Sheba announced.

The more consequential piece of the deal is what comes next: Sheba will integrate its own clinical protocols and policy documents directly into the platform, so the AI’s responses reflect how Sheba specifically practices medicine rather than general medical knowledge alone.

“We are not just implementing AI in medicine but building a fully AI-powered hospital,” said Dr. Eyal Zimlichman, chief innovation officer at Sheba and founder of ARC (Accelerate, Redesign, Collaborate), Sheba’s healthcare innovation and transformation arm. OpenAI’s models will not be trained on Sheba’s data, and patient information stays protected under through data isolation and audit protocols, i24NEWS reported.

Final medical decisions remain the sole responsibility of Sheba’s clinicians. Sheba will also get early access to OpenAI’s newest models through its API for research.

Banks Already Learned Generic AI Models Are Not Enough

Grounding a general-purpose model in an institution’s own rules is not new. Banks did this years earlier. A generic AI model that ignored a bank’s lending policies, fraud thresholds or compliance rules risked violating the regulations banks are required to follow.

The result is that AI vendors are now building directly into bank infrastructure instead of selling a standalone product. For example, Anthropic launched 10 financial-services AI agents this year for underwriting reviews, know-your-customer checks and compliance work, tools designed to plug into a bank’s existing risk and compliance systems, PYMNTS reported.

Experian built an Agent Operating System inside its Ascend Platform specifically to give lenders auditability and human oversight controls across every stage of an AI agent’s role in a lending decision, PYMNTS reported. In both cases, the value a bank pays for is not the underlying model. It is the layer that makes the model behave like this specific bank rather than any bank.

Cedars-Sinai Is Feeding Its Internal Workflows Into OpenEvidence

Hospitals face the same constraint banks faced. A general-purpose model reflects broad medical literature, not a hospital’s own treatment pathways and protocols. Cedars-Sinai made that gap explicit in May when it partnered with clinical AI platform OpenEvidence. The hospital said it plans to incorporate its own care pathways and best practices into the platform, so clinicians see medical literature alongside Cedars-Sinai-specific guidance, Newswise reported. “Medicine is not practiced in the abstract. It is practiced on individual patients with unique histories and complexities,” OpenEvidence CEO Daniel Nadler said.

Mount Sinai took the same approach two months earlier, becoming OpenEvidence’s first enterprise health system deal. The seven-hospital network gave physicians, nurses and pharmacists access to a platform pulling from clinical guidelines and peer-reviewed literature, Becker’s Hospital Review reported. “We are equipping every member of the care team with real-time access to rigorously sourced, evidence-based insights,” said Girish Nadkarni, Mount Sinai’s chief AI officer.

The underlying model is becoming commodity infrastructure. The proprietary rules layered on top are becoming the differentiator. Cedars-Sinai, Mount Sinai and now Sheba are each betting that a model trained on their own protocols will outperform a more capable model that isn’t.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.